Speaker Adaptation in Continuous Speech Recognition Using MLLR-Based MAP Estimation

A variety of methods are used for speaker adaptation in speech recognition. In some techniques, such as MAP estimation, only the models with available training data are updated. Hence, large amounts of training data are required in order to have significant recognition improvements. In some others, such as MLLR, where several general transformations are applied to model clusters, the results are desirable for small training data, but with increasing training data, the performance improvement reaches the saturation lvel. In this paper, a new approach is introduced that makes use of the advantages of both mentioned techniques to improve the recognition rate. Here, the models with available training data are trained using MAP while for those with insufficient training data, appropriate prior parameters for MAP estimation are found using MLLR. This technique has yielded better performance in comparison to either MAP or MLLR, in a system based on FARSDAT speech corpus.